Imagine you're the AI. Give yourself a solid minute to brainstorm ideas.

Here's my answer, as a non-superintelligent human: "see to it that the humans on top of the situation have a compelling financial interest in the systems not disconnecting".

In nuclear engineering, where safety is taken seriously, it's not enough to end the conversation at "the humans in charge can always simply shut down the reactor during a meltdown" or "a meltdown has never happened before, so we don't have to design safety systems before one does".